Blog

Five-agent LLM committees collapse to 1.2–1.8 effective independent voters. We ran extensive benchmarks and found a cheaper calibration-based fix.

A linear probe separates EvalAwareBench from WildChat with AUROC 0.986, but artifact controls and base rates tell a more sobering story.

A four-layer, domain-agnostic architecture for policy simulation using attribute-conditioned LM agents in a closed-loop game framework with co-adaptive mechanisms.